Insertion-deletion regulation prediction method and device based on epigenomics and medium

Abstract

The application discloses an indel regulation prediction method and device based on epigenomics and a medium, and the method comprises the following steps: obtaining tissue-specific non-coding indel mutation data and tissue-specific epigenomics data, calculating the enrichment of indel mutation sites and epigenetic modification sites between corresponding tissues, and constructing a balanced positive and negative sample set; the missing values in the genomic function annotation are filled by using the median filling method, and the filled genomic function annotation and the sequence category feature are spliced to serve as input features of the model; the best parameters of the model are found by using Bayesian optimization, and the model is trained and predicted; the tissue-specific epigenomics data is used to divide positive and negative samples, and the comprehensive genomic function annotation and the sequence category feature are combined to construct a tissue-specific indel regulation prediction model with excellent performance, so that the indel mutations with potential regulation functions can be screened more efficiently.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of machine learning and bioinformatics, and particularly relates to a method for predicting tissue-specific non-coding indel mutation regulation based on epigenomics. BACKGROUND

[0002] Since non-coding indel mutations do not encode proteins, it was previously believed that their impact on diseases was far less than that of coding region indel mutations. However, as research deepens, it has been found that non-coding indel mutations can still affect the role and influence of intervention regulatory elements (such as promoters, enhancers, etc.) in the process of gene expression, cause negative functional consequences (such as prohibiting inappropriate expression of genes) of gene expression, and become potential factors for the occurrence and development of diseases. However, biological experiments to study the regulatory function of non-coding indel mutations are time-consuming and labor-intensive, and in comparison, computational-based methods are efficient and low-cost, and can also provide a basis for further experimental verification. In view of this, a small number of non-coding indel mutation prediction tools have been proposed, which can be divided into two categories according to the scope of application, namely general-purpose tools and specific tools. Among them, general-purpose tools do not distinguish between point mutations and indel mutations, so the prediction effect on non-coding indel mutations is still insufficient, and the performance is not as good as specific tools. However, for specific tools, due to the lack of experimentally verified positive sample data of regulatory non-coding indel mutations, the data of existing tools only rely on statistical indicators to divide positive and negative samples, and cannot reflect the real biological phenomenon. The current prediction methods for the regulatory function of non-coding indel mutations are difficult to achieve satisfactory results.

[0003] How to solve the above problems, the need to propose a new technical means. SUMMARY

[0004] Based on the above problems, the present application proposes a method for predicting regulatory tissue-specific non-coding indel mutations based on epigenomics.

[0005] The technical problem to be solved by the present application is:

[0006] How to take appropriate methods to construct a model for effective prediction and classification of regulatory tissue-specific non-coding indel mutations for low-confidence data.

[0007] The purpose of the present application can be achieved by the following technical solutions:

[0008] A method for predicting tissue-specific non-coding mutations based on epigenomics, characterized in that:

[0009] Dataset construction step: The tissue-specific non-coding indel mutation data were obtained from the database GTEx, and four kinds of epigenomic data (DNase I hypersensitive sites (DHS), histone modification, enhancer RNA and DNase I footprinting data) were obtained from other public literatures. For each tissue, the enrichment of mutation sites with each epigenetic data site was calculated using the genomics tool Bedtools, and the mutations without site enrichment were assigned to the negative sample set, and the mutations with enrichment were assigned to the positive sample set. Then, according to the ranking of each mutation enrichment site in the positive sample set, when the positive samples are more than the negative samples, the same amount of positive samples as the negative samples are selected in descending order of the number of enrichment sites; if the negative samples are more than the positive samples, the negative samples are randomly sampled without replacement, so that a balanced positive and negative sample set is constructed for each tissue.

[0010] Feature extraction and processing step: According to the mutation information, the genomic function annotation of each mutation provided by the mutation annotation tool CADDv1.7 in the public literature is collected; then the mutation information is input into the epigenetic map prediction model Sei (so far the most comprehensive epigenetic prediction map model), and the epigenetic map contained in the sequence where the mutation is located is obtained using it, and further sequence class features are calculated. For genomic function annotation features, only genomic function annotations with less than 10% missing values are retained, and the remaining missing values are processed using the median imputation method. Then, using the feature fusion splicing method, the processed genomic function annotations and sequence class features are integrated together to make them the input features of the model, and the above process is performed for each tissue.

[0011] Model construction step: The best parameters of the model are found by using Bayesian optimization. On this basis, experiments are carried out on six common machine learning classifiers, including support vector machine (SVM), random forest (RF), logistic regression (LR), extreme gradient boosting (XGBoost) and adaptive boosting algorithm (AdaBoost). The feature matrix and label of each tissue are input into the above machine learning classifiers, and the best parameters of each classifier are found by Bayesian optimization to achieve the best performance of each model. For each classifier, the corresponding model is constructed and evaluated, and the performance of each classifier is compared, and finally the extreme gradient boosting (XGBoost) with the best average index in five-fold cross-validation is selected as the classifier used to construct the final model.

[0012] Model training step: A fine nested cross-validation strategy was used, including inner and outer loops. First, the whole dataset was randomly divided into training and test sets at a ratio of 8:2. The training set entered the inner loop and was further divided into training and validation sets at a ratio of 8:2. The training and validation sets were used to optimize the model parameters by Bayesian optimization to find the best parameter optimization model. The test set was used in the outer loop to verify the final performance of the model with the best parameters. For each tissue sample, the whole process was repeated 5 times, and the average of the 5 evaluation scores was taken as the final performance indicator of the model to ensure the stability and excellence of the model under different random seeds and different tissues.

[0013] Model prediction step: The random seed was fixed to ensure the reproducibility of the results. The test set randomly divided from the original dataset at a ratio of 8:2 was used to predict on each tissue model, and 5 prediction results (evaluation indicators) were obtained by randomly dividing 5 times. The average of the 5 prediction results was taken as the final evaluation indicator of the test set on the model. The final model was obtained by integrating the five-fold cross-validation model with the optimal parameters to obtain the regulatory probability of tissue-specific non-coding indel mutations and realize the prediction of tissue-specific non-coding indel mutation data. When external data wants to be predicted, the input features can be directly input into the model for prediction to obtain the regulatory probability.

[0014] Dataset construction step description:

[0015] Four epigenomic data were obtained from public literature: DNase I hypersensitive sites (DHS), histone modification, enhancer RNA, and DNase I footprinting (Consensus footprints) data without tissue / cell specificity. For each tissue, the enrichment of non-coding causal indel mutation sites with four epigenomic sites was calculated using the bioinformatics tool Bedtools. For a certain non-coding indel mutation, if there is no enrichment site on all four types of genomic data, then this mutation is included in the negative sample set; otherwise, there is an enrichment site on four types of genomic data (including one or more), then this mutation is included in the positive sample set.

[0016] After the preliminary sample screening is completed, the step of constructing a balanced positive and negative sample set is entered. First, focus on the positive sample set, calculate the enrichment site number of each mutation and epigenetic occurrence site and sort. If the positive samples are more than the negative samples, select the same amount of positive samples as the negative samples in descending order of enrichment site number, for example, if there are 500 negative samples and 800 positive samples, the first 500 positive samples are selected to form a balanced set with the negative samples; on the contrary, if the negative samples are more than the positive samples, use non-replacement random sampling to extract the same amount of samples from the more negative samples as the positive samples, for example, if there are 300 positive samples and 600 negative samples, extract 300 negative samples to construct a balanced set with the positive samples. In this way, 11 balanced sample datasets of tissues are constructed, including 11 tissues of adrenal gland, sigmoid colon, esophageal mucosa, left ventricle, liver, lung, skeletal muscle, ovary, pancreas, spleen and stomach.

[0017] In yet another aspect, the present application also discloses a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to make the processor execute the steps of the above method.

[0018] In yet another aspect, the present application also discloses a computer device, which comprises a memory and a processor, and the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the above method.

[0019] In summary, the application discloses an indel regulation prediction method and device based on epigenomics and a medium, the method comprising the following steps: (1) data set construction, obtaining tissue-specific non-coding indel mutation data and tissue-specific epigenomics data from a public database, calculating the enrichment of indel mutation sites and epigenetic modification sites between corresponding tissues, and constructing a balanced positive and negative sample set; (2) feature extraction and processing, filling in missing values in genomic function annotation by using the median filling method, and splicing the filled genomic function annotation and sequence class features as input features of the model; (3) model construction, finding the best parameters of the model by using Bayesian optimization, and performing experiments on common machine learning classifiers to select the classifier with the best performance; (4) model training, dividing the constructed tissue-specific data set into a test set, then dividing the remaining data set into a training set and a validation set, performing model evaluation and parameter optimization on the validation set, and obtaining the final tissue-specific indel mutation prediction model by integrating the nested five-fold cross-validation model under the optimal parameters on the training set; (5) model prediction, using the divided test set for prediction to obtain the regulatory probability. The method divides positive and negative samples by using tissue-specific epigenomics data, combines comprehensive genomic function annotation and sequence class features, and constructs a tissue-specific indel regulation prediction model with excellent performance, helping researchers to more efficiently screen indel mutations with potential regulatory functions and accelerate the in-depth understanding of gene regulation networks. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below in conjunction with the drawings of the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments of the present application.

[0021] Figure 1 is the overall framework flow schematic diagram of the present application. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below in conjunction with the drawings of the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments of the present application.

[0023] As Figure 1The tissue-specific non-coding indel mutation regulation prediction method based on epigenomics of the embodiment of the application is shown, and includes the following steps: (1) dataset construction, obtaining tissue-specific non-coding indel mutation data and tissue-specific epigenomics data from a public database, calculating the enrichment of indel mutation sites and epigenetic modification sites between corresponding tissues, and constructing a balanced positive and negative sample set; (2) feature extraction and processing, filling in missing values in genomic functional annotations using the median filling method, and splicing the filled genomic functional annotations and sequence class features as input features of the model; (3) model construction, using Bayesian optimization to find the best parameters of the model, and performing experiments on common machine learning classifiers to select the classifier with the best performance; (4) model training, dividing the constructed tissue-specific dataset into a test set, and then dividing the remaining dataset into a training set and a validation set, performing model evaluation and parameter optimization on the validation set, and obtaining the final tissue-specific indel mutation prediction model through nested five-fold cross-validation of the model with the optimal parameters on the training set; (5) model prediction, using the divided test set for prediction to obtain the regulatory probability.

[0024] Dataset construction, obtaining tissue-specific non-coding causal indel mutation data from the database GTEx, and obtaining epigenomics data from other public literature. The site enrichment is calculated, and the mutations without site enrichment in the corresponding tissue are classified into the negative sample set, and the mutations with enrichment are classified into the positive sample set. Then, according to the ranking of the enrichment sites in the positive sample set, when the number of positive samples is more than that of negative samples, the same number of positive samples as negative samples is selected in descending order of the number of enrichment sites; if the number of negative samples is more than that of positive samples, no-random random sampling is used for negative samples, so that a balanced positive and negative sample set is constructed for each tissue.

[0025] Feature extraction and processing, collecting genomic functional annotations provided by CADDv1.7 and sequence class features provided by the epigenetic map prediction model Sei. The genomic functional annotations with less than 10% missing values are screened out, and the remaining missing values are processed using the median filling method. Then, the processed genomic functional annotations and sequence class features are integrated together by feature fusion splicing, so as to become the input features of the model.

[0026] Model construction, Bayesian optimization is used to find the best parameters of the model. On this basis, experiments are carried out on six common machine learning classifiers, including support vector machine (SVM), random forest (RF), logistic regression (LR), extreme gradient boosting (XGBoost) and adaptive boosting algorithm (AdaBoost). For each classifier, the corresponding model is constructed and evaluated, and the performance of each classifier is compared. Finally, the extreme gradient boosting (XGBoost) with the best performance is selected as the classifier for building the final model.

[0027] Model training, a fine nested cross-validation strategy is used, including inner and outer loops. First, the entire dataset is randomly divided into training and test sets in the ratio of 8:2. The training set enters the inner loop and is further divided into 8:2 to form the validation set. Bayesian optimization is used to optimize the model using the training and validation sets. The test set is used in the outer loop to verify the final performance of the model with the best parameters. For each tissue sample, the entire process is repeated 5 times, and the average of the 5 evaluation scores is taken as the final performance indicator of the model to ensure the stability and excellence of the model in different tissues.

[0028] Model prediction, the test set randomly divided in the ratio of 8:2 is used to predict the regulatory probability of tissue-specific non-coding indel mutations on each tissue model, achieving the prediction of tissue-specific non-coding indel mutation data.

[0029] Among them, the data set construction method is:

[0030] Four epigenomic data sets are obtained from public literature: DNase I hypersensitive sites (DHS), histone modification, enhancer RNA, and DNase I footprinting (Consensus footprints) data without tissue / cell specificity. For each tissue, the enrichment of non-coding causal indel mutation sites and four epigenomic sites is calculated using the bioinformatics tool Bedtools. For a non-coding indel mutation, if there is no enrichment site on all four epigenomic data, then this mutation is included in the negative sample set; otherwise, there is an enrichment site on four epigenomic data (including one or more), then this mutation is included in the positive sample set.

[0031] After the preliminary sample screening is completed, the step of constructing a balanced positive and negative sample set is entered. First, focus on the positive sample set, calculate the enrichment site number of each mutation and epigenetic site, and sort them. If the positive samples are more than the negative samples, select the same number of positive samples as the negative samples in descending order of enrichment site number. For example, if there are 500 negative samples and 800 positive samples, the first 500 positive samples are selected to form a balanced set with the negative samples. Conversely, if the negative samples are more than the positive samples, use non-replacement random sampling to extract the same number of samples from the more negative samples as the positive samples. For example, if there are 300 positive samples and 600 negative samples, 300 negative samples are extracted to construct a balanced set with the positive samples. In this way, 11 balanced sample datasets of tissues are constructed, including 11 tissues of adrenal gland, sigmoid colon, esophageal mucosa, left ventricle, liver, lung, skeletal muscle, ovary, pancreas, spleen and stomach.

[0032] For each tissue, a tissue-specific model is constructed separately, and the final model is obtained by integrating the five-fold cross-validation model under the optimal parameters.

[0033] The feature splicing step is described, and the formula for splicing the same number of genomic function annotations and sequence category features of the samples is as follows:

[0034]

[0035] In the formula, a ij represents an mxn1 matrix A, where i=1, 2, …, m and j=1, 2, …, n1; b ij represents an mxn2 matrix B, where i=1, 2, …, m and j=1, 2, …, n2. The transverse splicing (splicing by column) obtains a matrix C, and the element c ij can be obtained from the above formula.

[0036] The following is a specific description:

[0037] The method for predicting regulatory tissue-specific non-coding indel mutations based on epigenomics according to an embodiment of the present application comprises:

[0038] (1) Data set construction: four kinds of epigenomics data and non-coding indel mutation data of 11 tissues are collected. For each tissue, the Bedtools tool is used to calculate the overlap and enrichment of mutation sites with four kinds of epigenomics data sites. The mutations without overlap and enrichment are included in the negative samples, and the rest are included in the positive samples. Then, a balanced positive and negative sample dataset is constructed by using a non-replacement random sampling method.

[0039] (2) Feature extraction and preprocessing, based on the model of mutation information published in the literature, two sources of feature input are extracted from genomic function annotation and sequence category features. In the feature preprocessing step, the missing value filling algorithm is used to fill the missing values of the genomic function annotation, and the splicing method in the feature fusion field is used to splice with the sequence category features.

[0040] (3) Model construction, using nested cross-validation and Bayesian optimization method to construct five-fold cross-validation models of multiple different machine learning classifiers. Then select the model of different classifiers based on performance evaluation, based on the optimal in 11 organizations to get the final optimal model using XGBoost. The input of the model is the spliced features after preprocessing, the classifier of the model is the machine learning algorithm XGBoost, and the output of the model is the prediction probability (i.e. the probability of each insertion-deletion mutation with regulatory function).

[0041] (4) Model training, the data set of 11 organizations is divided into training set and test set respectively, the training set is used to train the model, the validation set is used to evaluate and optimize the model parameters, and the best parameters are fixed after finding the best parameters, the test set is used for further evaluation, and the five-fold cross-validation model with optimal parameters is also optimized to achieve the optimal performance index, and this optimal model is used as the final 11 organization-specific prediction model;

[0042] (5) Model prediction, the test set is based on the five-fold cross-validation random division of the original data set, and the constructed model is used to perform five-fold cross-validation prediction on the divided test set to obtain its regulatory probability.

[0043] Overall, the present application how to construct a model with excellent performance in the case of low confidence data to predict and classify regulatory non-coding insertion-deletion mutations, which contributes to the subsequent discovery of regulatory tissue-specific non-coding insertion-deletion mutations and the study of tissue-specific expression gene regulatory network.

[0044] Embodiment:

[0045] The data set of the experiment is composed of 11 different tissue data sets: 11 tissue-specific non-coding indel data sets containing adrenal gland, sigmoid colon, esophageal mucosa, left ventricle, liver, lung, skeletal muscle, ovary, pancreas, spleen, and stomach. Based on four kinds of epigenomics data (DNase I hypersensitive sites, histone modification, enhancer RNA, and DNase I footprinting data without tissue / cell specificity), the present application calculates the enrichment site of the non-coding indel mutation source data derived from the eighth edition of the GTEx database, and obtains 11 tissue-specific balanced data sets by non-replacement random sampling. DNase I hypersensitive sites (DHS) and histone modification data are derived from the public literature [LU H, MA L, QUAN C, et al. RegVar: Tissue-Specific Prioritization of Non-Coding Regulatory Variants [J]. Genomics, Proteomics & Bioinformatics, 2021, 21(2): 385-95.], enhancer RNA data and DNase I footprinting data are derived from the public literature [BIDDIE S C, WEYKOPF G, HIRD E F, et al. DNA-binding factor footprints and enhancer RNAs identify functional non-coding genetic variants [J]. Genome Biol, 2024, 25(1): 208.].

[0046] According to the mutation information, the genomic function annotation of each mutation provided by the mutation annotation tool CADDv1.7 provided in the public literature is collected; then the mutation information is input into the epigenetic map prediction model Sei (the most comprehensive epigenetic prediction map model so far), and the epigenetic map contained in the sequence where the mutation is located is obtained using the model, and sequence category features are further calculated. The genomic function annotation is derived from the annotation information provided in the public literature [SCHUBACH M, MAASS T, NAZARETYAN L, et al. CADD v1.7: using protein language models, regulatory CNNs and other nucleotide-level scores to improve genome-wide variant predictions [J]. Nucleic Acids Research, 2024, 52 (D1): D1143-D54.], and the sequence category features are derived from the deep learning model extraction provided in the prediction model literature [CHEN K M, WONG A K, TROYANSKAYA O G, et al. A sequence-based global map of regulatory activity for deciphering human genetics [J]. Nat Genet, 2022, 54 (7): 940-9.]. After feature extraction and missing value processing, the present application first uses a splicing method of feature fusion to construct the splicing input features of the model, and then selects the optimal performance model by constructing different machine learning classifier models. The present application uses nested five-fold cross-validation and Bayesian parameter optimization to select and optimize the model, and finally integrates the five-fold cross-validation model under the optimal parameters as the final model, and names it RegTIND (Regulatory Tissue-specific nc-INDel).

[0047] The present application uses random partition test set to comprehensively evaluate the performance of the model. Wherein the test set of each tissue accounts for 20% of the original data set, and the composition of the entire data set used in the experiment is shown in Table 1;

[0048] Table 1 Data set composition

[0049]

[0050]

[0051] The evaluation indicators used by the present application include recall (Recall), also known as sensitivity (Sensitivity, SEN), precision (Precision, PRE), accuracy (Accuracy, ACC) and F1 score (F1 score, F1), and their calculation methods are as follows:

[0052]

[0053] Wherein, TP (True positive) represents the number of true positives, that is, the number of true regulatory non-coding indel mutations correctly predicted as regulatory non-coding indel mutations, TN (True negative) represents the number of true negatives, that is, the number of true non-regulatory non-coding indel mutations correctly predicted as non-regulatory non-coding indel mutations, FP (False positive) is the number of false positives, that is, the number of regulatory non-coding indel mutations originally predicted as non-regulatory non-coding indel mutations, and FN (False negative) is the number of false negatives, that is, the number of non-regulatory non-coding indel mutations originally predicted as regulatory non-coding indel mutations. In addition, AUC and AUPR are used in the present patent to measure the overall performance of the model. Generally, the four indicators given in the above formula will be affected by the threshold value, that is, greater than or equal to the threshold value is predicted as a positive sample, and less than the threshold value is considered as a negative sample, and the default value of the threshold value is 0.5, but it can be manually adjusted. AUC and AUPR are not affected by the threshold value, and the range is between 0 and 1, the closer to 1, the better the overall performance of the model, so it is often considered to be a more important evaluation indicator.

[0054] Table 2 Comparison of five-fold cross-validation average values of RegTIND and other methods based on 11 tissue test sets

[0055]

[0056] Table 3 Comparison of five-fold cross-validation of RegTIND and other methods based on 11 tissue test sets

[0057]

[0058]

[0059] In order to verify the superiority of the model of the present application, the present application selects some excellent tools for comparison, including: CADD and TIVAN-indel. Among them, TIVAN-indel belongs to a specific tool, and CADD belongs to a broad-spectrum tool. Table 2 and Table 3 fully show the performance comparison of RegTIND with the other two tools under six evaluation indexes (SEN, PRE, F1, ACC, AUC and AUPR). In terms of sensitivity, the average value of RegTIND in 11 tissue test sets is 0.951, which is higher than that of other methods (the sensitivity of other methods in the test set is between 0.020-0.707), which shows that the method of the present application has good recognition ability and coverage ability. In addition, the AUC and AUPR of RegTIND on the five-fold cross-validation test set are higher than those of other methods, and the average AUC result of RegTIND on the five-fold cross-validation test set in 11 tissues is 0.144-0.440 higher than that of other tools, and the average AUPR result of RegTIND on the five-fold cross-validation test set in 11 tissues is 0.113-0.439 higher than that of other tools. At the same time, from the comparison results of the three tools in the 11 tissues shown in Table 3, it can be seen that RegTIND shows the best performance in all tissues, which shows that the method of the present application has relatively accurate prediction ability in all 11 tissues. The above analysis shows that RegTIND has strong generalization ability.

[0060] In yet another aspect, the present application also discloses a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to make the processor execute the steps of the above method.

[0061] In yet another aspect, the present application also discloses a computer device, which comprises a memory and a processor, and the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the above method.

[0062] In yet another embodiment provided in the present application, a computer program product containing instructions is also provided, which makes a computer execute the above-mentioned any one of the epigenetic genomics-based indel regulation prediction methods when running on the computer.

[0063] It can be understood that the system, device and storage medium provided by the embodiments of the present application correspond to the method provided by the embodiments of the present application, and the explanation, examples and beneficial effects of the related contents can be referred to the corresponding part in the above method.

[0064] In the embodiments described above, all or some of the steps can be implemented by hardware, software, firmware or any combination thereof. When implemented by software, all or some of the steps can be implemented in the form of one or more computer programs or program elements. The computer programs reside (at least temporarily) in a memory of a computer during execution. The memory can be a RAM memory, a flash memory, a ROM memory, an EPROM memory, or any other suitable memory. The memory can be integral to or separate from the computer. The computer programs can be written in any suitable programming language, such as C, C++, Java, Visual Basic, etc. The computer programs can be written in assembly or machine language, if desired. The computer programs can be distributed over network coupled file servers, or can be distributed by any other suitable means.

[0065] It is to be understood that the terminology “first”, “second”, etc. is used in this document only to distinguish one entity or operation from another, and does not necessarily imply or suggest any actual relationship or order between or among such entities or operations. Also, the terms “comprises”, “comprising”, or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by “comprises...” does not, without more restrictions, exclude the existence of additional identical elements in a process, method, article, or apparatus that includes the stated element.

[0066] Each of the embodiments described in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other. Each of the embodiments focuses on the difference from other embodiments. In particular, for the system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the description of the method embodiments.

[0067] The above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting insertion and deletion regulation based on epigenetics, characterized in that, Includes the following steps, Dataset construction: Obtain tissue-specific non-coding insertion and deletion mutation data and tissue-specific epigenetics data from public databases, calculate the enrichment of insertion and deletion mutation sites and epigenetic modification sites between corresponding tissues, and construct a balanced set of positive and negative samples. Feature extraction and processing: The median imputation method is used to fill missing values ​​in the genome functional annotation, and the imputed genome functional annotation is concatenated with the sequence category features as the input features of the model; Model building: Bayesian optimization is used to find the optimal parameters of the model, and experiments are conducted on common machine learning classifiers to select the classifier with the best performance; Model training: The constructed tissue-specific dataset is divided into a test set, and the remaining dataset is divided into a training set and a validation set. Model evaluation and parameter optimization are performed on the validation set. The final tissue-specific insertion and deletion mutation prediction model is obtained by integrating the nested five-fold cross-validation model with the optimal parameters on the training set. Model prediction: The tissue-specific insertion and deletion mutation prediction model is used to make predictions based on the partitioned test set to obtain its regulatory probability; Feature extraction specifically includes, Genomic functional annotations for each mutation were collected using CADDv1.7, a mutation annotation tool provided in publicly available literature, based on mutation information. Next, the mutation information is input into the epigenetic map prediction model Sei to obtain the epigenetic map contained in the sequence where the mutation is located, and further calculate and extract the sequence category features; Feature extraction is followed by processing steps. Based on the characteristics of genome functional annotation, only genome functional annotations with missing values ​​of less than 10% are retained, and the remaining missing values ​​are processed using the median imputation method; Subsequently, a feature fusion splicing method was used to integrate the processed genome functional annotations with sequence category features, which became the input features of the model.

2. The method for predicting insertion and deletion regulation based on epigenetics according to claim 1, characterized in that: The steps of the splicing method using feature fusion include splicing genome functional annotations and sequence category features with the same number of samples. The formula is as follows: In the formula, Represents m× Matrix A, where and ; Represents m× The matrix B, where and ; Horizontal concatenation, that is, concatenating columns, yields matrix C, whose elements It can be obtained from the above formula.

3. The method for predicting insertion and deletion regulation based on epigenetics according to claim 1, characterized in that: Dataset construction included obtaining tissue-specific non-coding causal insertion and deletion mutation data from the GTEx database and epigenetics data from other publicly available literature; The site enrichment is calculated, and mutations with no site enrichment in the corresponding tissue are included in the negative sample set, while those with enrichment are included in the positive sample set. Subsequently, based on the sorting of each mutation enrichment site in the positive sample set, when there are more positive samples than negative samples, positive samples of equal number to negative samples are selected in descending order of the number of enrichment sites; if there are more negative samples than positive samples, random sampling without replacement is used for negative samples, thereby constructing a balanced set of positive and negative samples in each tissue.

4. The method for predicting insertion and deletion regulation based on epigenetics according to claim 1, characterized in that: The model building steps include finding the optimal parameters of the model using Bayesian optimization, and then conducting experiments on six common machine learning classifiers: support vector machine, random forest, logistic regression, extreme gradient boosting, and adaptive boosting algorithm. The feature matrix and labels of each organization are input into the machine learning classifiers mentioned above, and Bayesian optimization is performed to find the best parameters for each classifier in order to achieve the best performance of each model. For each classifier, a corresponding model was built and evaluated. The performance of each classifier was compared, and the extreme gradient booster that achieved the best average index in five-fold cross-validation was finally selected as the classifier used to build the final model.

5. The method for predicting insertion and deletion regulation based on epigenetics according to claim 1, characterized in that: The model training steps include employing a sophisticated nested cross-validation strategy, including inner and outer loops; First, the entire dataset is randomly divided into a training set and a test set in a ratio of 8:

2. The training set is then subdivided into a validation set in an inner loop in a ratio of 8:

2. The training set and validation set are then used for Bayesian optimization to tune the parameters and find the optimal parameters to optimize the model. The test set is kept in the outer loop to verify the final performance of the model with the best parameters. For each tissue sample, the entire process was repeated 5 times, and the average of the 5 evaluation scores was taken as the final performance index of the model, ensuring that the model is stable and excellent under different random seeds and different tissues.

6. The method for predicting insertion and deletion regulation based on epigenetics according to claim 1, characterized in that: The model prediction steps include fixing the random seed to ensure the reproducibility of the results, using a test set randomly divided from the original dataset in an 8:2 ratio on each organizational model for prediction, randomly dividing the dataset 5 times to obtain 5 prediction results, and using the average of these 5 prediction results as the final evaluation metric for the test set on the model. The final model was obtained by integrating a five-fold cross-validation model with optimal parameters to obtain the regulatory probability of tissue-specific non-coding insertion and deletion mutations, thereby enabling the prediction of tissue-specific non-coding insertion and deletion mutation data. When there is external data that needs to be predicted, the input features can be directly input into the model for prediction to obtain the regulatory probability.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, the processor performs the steps of the method as described in any one of claims 1 to 6.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, the processor performs the steps of the method as described in any one of claims 1 to 6.

Images